Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * trace_hwlat.c - A simple Hardware Latency detector.
4 *
5 * Use this tracer to detect large system latencies induced by the behavior of
6 * certain underlying system hardware or firmware, independent of Linux itself.
7 * The code was developed originally to detect the presence of SMIs on Intel
8 * and AMD systems, although there is no dependency upon x86 herein.
9 *
10 * The classical example usage of this tracer is in detecting the presence of
11 * SMIs or System Management Interrupts on Intel and AMD systems. An SMI is a
12 * somewhat special form of hardware interrupt spawned from earlier CPU debug
13 * modes in which the (BIOS/EFI/etc.) firmware arranges for the South Bridge
14 * LPC (or other device) to generate a special interrupt under certain
15 * circumstances, for example, upon expiration of a special SMI timer device,
16 * due to certain external thermal readings, on certain I/O address accesses,
17 * and other situations. An SMI hits a special CPU pin, triggers a special
18 * SMI mode (complete with special memory map), and the OS is unaware.
19 *
20 * Although certain hardware-inducing latencies are necessary (for example,
21 * a modern system often requires an SMI handler for correct thermal control
22 * and remote management) they can wreak havoc upon any OS-level performance
23 * guarantees toward low-latency, especially when the OS is not even made
24 * aware of the presence of these interrupts. For this reason, we need a
25 * somewhat brute force mechanism to detect these interrupts. In this case,
26 * we do it by hogging all of the CPU(s) for configurable timer intervals,
27 * sampling the built-in CPU timer, looking for discontiguous readings.
28 *
29 * WARNING: This implementation necessarily introduces latencies. Therefore,
30 * you should NEVER use this tracer while running in a production
31 * environment requiring any kind of low-latency performance
32 * guarantee(s).
33 *
34 * Copyright (C) 2008-2009 Jon Masters, Red Hat, Inc. <jcm@redhat.com>
35 * Copyright (C) 2013-2016 Steven Rostedt, Red Hat, Inc. <srostedt@redhat.com>
36 *
37 * Includes useful feedback from Clark Williams <williams@redhat.com>
38 *
39 */
40#include <linux/kthread.h>
41#include <linux/tracefs.h>
42#include <linux/uaccess.h>
43#include <linux/cpumask.h>
44#include <linux/delay.h>
45#include <linux/sched/clock.h>
46#include "trace.h"
47
48static struct trace_array *hwlat_trace;
49
50#define U64STR_SIZE 22 /* 20 digits max */
51
52#define BANNER "hwlat_detector: "
53#define DEFAULT_SAMPLE_WINDOW 1000000 /* 1s */
54#define DEFAULT_SAMPLE_WIDTH 500000 /* 0.5s */
55#define DEFAULT_LAT_THRESHOLD 10 /* 10us */
56
57static struct dentry *hwlat_sample_width; /* sample width us */
58static struct dentry *hwlat_sample_window; /* sample window us */
59static struct dentry *hwlat_thread_mode; /* hwlat thread mode */
60
61enum {
62 MODE_NONE = 0,
63 MODE_ROUND_ROBIN,
64 MODE_PER_CPU,
65 MODE_MAX
66};
67static char *thread_mode_str[] = { "none", "round-robin", "per-cpu" };
68
69/* Save the previous tracing_thresh value */
70static unsigned long save_tracing_thresh;
71
72/* runtime kthread data */
73struct hwlat_kthread_data {
74 struct task_struct *kthread;
75 /* NMI timestamp counters */
76 u64 nmi_ts_start;
77 u64 nmi_total_ts;
78 int nmi_count;
79 int nmi_cpu;
80};
81
82static struct hwlat_kthread_data hwlat_single_cpu_data;
83static DEFINE_PER_CPU(struct hwlat_kthread_data, hwlat_per_cpu_data);
84
85/* Tells NMIs to call back to the hwlat tracer to record timestamps */
86bool trace_hwlat_callback_enabled;
87
88/* If the user changed threshold, remember it */
89static u64 last_tracing_thresh = DEFAULT_LAT_THRESHOLD * NSEC_PER_USEC;
90
91/* Individual latency samples are stored here when detected. */
92struct hwlat_sample {
93 u64 seqnum; /* unique sequence */
94 u64 duration; /* delta */
95 u64 outer_duration; /* delta (outer loop) */
96 u64 nmi_total_ts; /* Total time spent in NMIs */
97 struct timespec64 timestamp; /* wall time */
98 int nmi_count; /* # NMIs during this sample */
99 int count; /* # of iterations over thresh */
100};
101
102/* keep the global state somewhere. */
103static struct hwlat_data {
104
105 struct mutex lock; /* protect changes */
106
107 u64 count; /* total since reset */
108
109 u64 sample_window; /* total sampling window (on+off) */
110 u64 sample_width; /* active sampling portion of window */
111
112 int thread_mode; /* thread mode */
113
114} hwlat_data = {
115 .sample_window = DEFAULT_SAMPLE_WINDOW,
116 .sample_width = DEFAULT_SAMPLE_WIDTH,
117 .thread_mode = MODE_ROUND_ROBIN
118};
119
120static struct hwlat_kthread_data *get_cpu_data(void)
121{
122 if (hwlat_data.thread_mode == MODE_PER_CPU)
123 return this_cpu_ptr(&hwlat_per_cpu_data);
124 else
125 return &hwlat_single_cpu_data;
126}
127
128static bool hwlat_busy;
129
130static void trace_hwlat_sample(struct hwlat_sample *sample)
131{
132 struct trace_array *tr = hwlat_trace;
133 struct trace_event_call *call = &event_hwlat;
134 struct trace_buffer *buffer = tr->array_buffer.buffer;
135 struct ring_buffer_event *event;
136 struct hwlat_entry *entry;
137
138 event = trace_buffer_lock_reserve(buffer, TRACE_HWLAT, sizeof(*entry),
139 tracing_gen_ctx());
140 if (!event)
141 return;
142 entry = ring_buffer_event_data(event);
143 entry->seqnum = sample->seqnum;
144 entry->duration = sample->duration;
145 entry->outer_duration = sample->outer_duration;
146 entry->timestamp = sample->timestamp;
147 entry->nmi_total_ts = sample->nmi_total_ts;
148 entry->nmi_count = sample->nmi_count;
149 entry->count = sample->count;
150
151 if (!call_filter_check_discard(call, entry, buffer, event))
152 trace_buffer_unlock_commit_nostack(buffer, event);
153}
154
155/* Macros to encapsulate the time capturing infrastructure */
156#define time_type u64
157#define time_get() trace_clock_local()
158#define time_to_us(x) div_u64(x, 1000)
159#define time_sub(a, b) ((a) - (b))
160#define init_time(a, b) (a = b)
161#define time_u64(a) a
162
163void trace_hwlat_callback(bool enter)
164{
165 struct hwlat_kthread_data *kdata = get_cpu_data();
166
167 if (!kdata->kthread)
168 return;
169
170 /*
171 * Currently trace_clock_local() calls sched_clock() and the
172 * generic version is not NMI safe.
173 */
174 if (!IS_ENABLED(CONFIG_GENERIC_SCHED_CLOCK)) {
175 if (enter)
176 kdata->nmi_ts_start = time_get();
177 else
178 kdata->nmi_total_ts += time_get() - kdata->nmi_ts_start;
179 }
180
181 if (enter)
182 kdata->nmi_count++;
183}
184
185/*
186 * hwlat_err - report a hwlat error.
187 */
188#define hwlat_err(msg) ({ \
189 struct trace_array *tr = hwlat_trace; \
190 \
191 trace_array_printk_buf(tr->array_buffer.buffer, _THIS_IP_, msg); \
192})
193
194/**
195 * get_sample - sample the CPU TSC and look for likely hardware latencies
196 *
197 * Used to repeatedly capture the CPU TSC (or similar), looking for potential
198 * hardware-induced latency. Called with interrupts disabled and with
199 * hwlat_data.lock held.
200 */
201static int get_sample(void)
202{
203 struct hwlat_kthread_data *kdata = get_cpu_data();
204 struct trace_array *tr = hwlat_trace;
205 struct hwlat_sample s;
206 time_type start, t1, t2, last_t2;
207 s64 diff, outer_diff, total, last_total = 0;
208 u64 sample = 0;
209 u64 thresh = tracing_thresh;
210 u64 outer_sample = 0;
211 int ret = -1;
212 unsigned int count = 0;
213
214 do_div(thresh, NSEC_PER_USEC); /* modifies interval value */
215
216 kdata->nmi_total_ts = 0;
217 kdata->nmi_count = 0;
218 /* Make sure NMIs see this first */
219 barrier();
220
221 trace_hwlat_callback_enabled = true;
222
223 init_time(last_t2, 0);
224 start = time_get(); /* start timestamp */
225 outer_diff = 0;
226
227 do {
228
229 t1 = time_get(); /* we'll look for a discontinuity */
230 t2 = time_get();
231
232 if (time_u64(last_t2)) {
233 /* Check the delta from outer loop (t2 to next t1) */
234 outer_diff = time_to_us(time_sub(t1, last_t2));
235 /* This shouldn't happen */
236 if (outer_diff < 0) {
237 hwlat_err(BANNER "time running backwards\n");
238 goto out;
239 }
240 if (outer_diff > outer_sample)
241 outer_sample = outer_diff;
242 }
243 last_t2 = t2;
244
245 total = time_to_us(time_sub(t2, start)); /* sample width */
246
247 /* Check for possible overflows */
248 if (total < last_total) {
249 hwlat_err("Time total overflowed\n");
250 break;
251 }
252 last_total = total;
253
254 /* This checks the inner loop (t1 to t2) */
255 diff = time_to_us(time_sub(t2, t1)); /* current diff */
256
257 if (diff > thresh || outer_diff > thresh) {
258 if (!count)
259 ktime_get_real_ts64(&s.timestamp);
260 count++;
261 }
262
263 /* This shouldn't happen */
264 if (diff < 0) {
265 hwlat_err(BANNER "time running backwards\n");
266 goto out;
267 }
268
269 if (diff > sample)
270 sample = diff; /* only want highest value */
271
272 } while (total <= hwlat_data.sample_width);
273
274 barrier(); /* finish the above in the view for NMIs */
275 trace_hwlat_callback_enabled = false;
276 barrier(); /* Make sure nmi_total_ts is no longer updated */
277
278 ret = 0;
279
280 /* If we exceed the threshold value, we have found a hardware latency */
281 if (sample > thresh || outer_sample > thresh) {
282 u64 latency;
283
284 ret = 1;
285
286 /* We read in microseconds */
287 if (kdata->nmi_total_ts)
288 do_div(kdata->nmi_total_ts, NSEC_PER_USEC);
289
290 hwlat_data.count++;
291 s.seqnum = hwlat_data.count;
292 s.duration = sample;
293 s.outer_duration = outer_sample;
294 s.nmi_total_ts = kdata->nmi_total_ts;
295 s.nmi_count = kdata->nmi_count;
296 s.count = count;
297 trace_hwlat_sample(&s);
298
299 latency = max(sample, outer_sample);
300
301 /* Keep a running maximum ever recorded hardware latency */
302 if (latency > tr->max_latency) {
303 tr->max_latency = latency;
304 latency_fsnotify(tr);
305 }
306 }
307
308out:
309 return ret;
310}
311
312static struct cpumask save_cpumask;
313
314static void move_to_next_cpu(void)
315{
316 struct cpumask *current_mask = &save_cpumask;
317 struct trace_array *tr = hwlat_trace;
318 int next_cpu;
319
320 /*
321 * If for some reason the user modifies the CPU affinity
322 * of this thread, then stop migrating for the duration
323 * of the current test.
324 */
325 if (!cpumask_equal(current_mask, current->cpus_ptr))
326 goto change_mode;
327
328 cpus_read_lock();
329 cpumask_and(current_mask, cpu_online_mask, tr->tracing_cpumask);
330 next_cpu = cpumask_next(raw_smp_processor_id(), current_mask);
331 cpus_read_unlock();
332
333 if (next_cpu >= nr_cpu_ids)
334 next_cpu = cpumask_first(current_mask);
335
336 if (next_cpu >= nr_cpu_ids) /* Shouldn't happen! */
337 goto change_mode;
338
339 cpumask_clear(current_mask);
340 cpumask_set_cpu(next_cpu, current_mask);
341
342 set_cpus_allowed_ptr(current, current_mask);
343 return;
344
345 change_mode:
346 hwlat_data.thread_mode = MODE_NONE;
347 pr_info(BANNER "cpumask changed while in round-robin mode, switching to mode none\n");
348}
349
350/*
351 * kthread_fn - The CPU time sampling/hardware latency detection kernel thread
352 *
353 * Used to periodically sample the CPU TSC via a call to get_sample. We
354 * disable interrupts, which does (intentionally) introduce latency since we
355 * need to ensure nothing else might be running (and thus preempting).
356 * Obviously this should never be used in production environments.
357 *
358 * Executes one loop interaction on each CPU in tracing_cpumask sysfs file.
359 */
360static int kthread_fn(void *data)
361{
362 u64 interval;
363
364 while (!kthread_should_stop()) {
365
366 if (hwlat_data.thread_mode == MODE_ROUND_ROBIN)
367 move_to_next_cpu();
368
369 local_irq_disable();
370 get_sample();
371 local_irq_enable();
372
373 mutex_lock(&hwlat_data.lock);
374 interval = hwlat_data.sample_window - hwlat_data.sample_width;
375 mutex_unlock(&hwlat_data.lock);
376
377 do_div(interval, USEC_PER_MSEC); /* modifies interval value */
378
379 /* Always sleep for at least 1ms */
380 if (interval < 1)
381 interval = 1;
382
383 if (msleep_interruptible(interval))
384 break;
385 }
386
387 return 0;
388}
389
390/*
391 * stop_stop_kthread - Inform the hardware latency sampling/detector kthread to stop
392 *
393 * This kicks the running hardware latency sampling/detector kernel thread and
394 * tells it to stop sampling now. Use this on unload and at system shutdown.
395 */
396static void stop_single_kthread(void)
397{
398 struct hwlat_kthread_data *kdata = get_cpu_data();
399 struct task_struct *kthread;
400
401 cpus_read_lock();
402 kthread = kdata->kthread;
403
404 if (!kthread)
405 goto out_put_cpus;
406
407 kthread_stop(kthread);
408 kdata->kthread = NULL;
409
410out_put_cpus:
411 cpus_read_unlock();
412}
413
414
415/*
416 * start_single_kthread - Kick off the hardware latency sampling/detector kthread
417 *
418 * This starts the kernel thread that will sit and sample the CPU timestamp
419 * counter (TSC or similar) and look for potential hardware latencies.
420 */
421static int start_single_kthread(struct trace_array *tr)
422{
423 struct hwlat_kthread_data *kdata = get_cpu_data();
424 struct cpumask *current_mask = &save_cpumask;
425 struct task_struct *kthread;
426 int next_cpu;
427
428 cpus_read_lock();
429 if (kdata->kthread)
430 goto out_put_cpus;
431
432 kthread = kthread_create(kthread_fn, NULL, "hwlatd");
433 if (IS_ERR(kthread)) {
434 pr_err(BANNER "could not start sampling thread\n");
435 cpus_read_unlock();
436 return -ENOMEM;
437 }
438
439 /* Just pick the first CPU on first iteration */
440 cpumask_and(current_mask, cpu_online_mask, tr->tracing_cpumask);
441
442 if (hwlat_data.thread_mode == MODE_ROUND_ROBIN) {
443 next_cpu = cpumask_first(current_mask);
444 cpumask_clear(current_mask);
445 cpumask_set_cpu(next_cpu, current_mask);
446
447 }
448
449 set_cpus_allowed_ptr(kthread, current_mask);
450
451 kdata->kthread = kthread;
452 wake_up_process(kthread);
453
454out_put_cpus:
455 cpus_read_unlock();
456 return 0;
457}
458
459/*
460 * stop_cpu_kthread - Stop a hwlat cpu kthread
461 */
462static void stop_cpu_kthread(unsigned int cpu)
463{
464 struct task_struct *kthread;
465
466 kthread = per_cpu(hwlat_per_cpu_data, cpu).kthread;
467 if (kthread)
468 kthread_stop(kthread);
469 per_cpu(hwlat_per_cpu_data, cpu).kthread = NULL;
470}
471
472/*
473 * stop_per_cpu_kthreads - Inform the hardware latency sampling/detector kthread to stop
474 *
475 * This kicks the running hardware latency sampling/detector kernel threads and
476 * tells it to stop sampling now. Use this on unload and at system shutdown.
477 */
478static void stop_per_cpu_kthreads(void)
479{
480 unsigned int cpu;
481
482 cpus_read_lock();
483 for_each_online_cpu(cpu)
484 stop_cpu_kthread(cpu);
485 cpus_read_unlock();
486}
487
488/*
489 * start_cpu_kthread - Start a hwlat cpu kthread
490 */
491static int start_cpu_kthread(unsigned int cpu)
492{
493 struct task_struct *kthread;
494
495 /* Do not start a new hwlatd thread if it is already running */
496 if (per_cpu(hwlat_per_cpu_data, cpu).kthread)
497 return 0;
498
499 kthread = kthread_run_on_cpu(kthread_fn, NULL, cpu, "hwlatd/%u");
500 if (IS_ERR(kthread)) {
501 pr_err(BANNER "could not start sampling thread\n");
502 return -ENOMEM;
503 }
504
505 per_cpu(hwlat_per_cpu_data, cpu).kthread = kthread;
506
507 return 0;
508}
509
510#ifdef CONFIG_HOTPLUG_CPU
511static void hwlat_hotplug_workfn(struct work_struct *dummy)
512{
513 struct trace_array *tr = hwlat_trace;
514 unsigned int cpu = smp_processor_id();
515
516 mutex_lock(&trace_types_lock);
517 mutex_lock(&hwlat_data.lock);
518 cpus_read_lock();
519
520 if (!hwlat_busy || hwlat_data.thread_mode != MODE_PER_CPU)
521 goto out_unlock;
522
523 if (!cpumask_test_cpu(cpu, tr->tracing_cpumask))
524 goto out_unlock;
525
526 start_cpu_kthread(cpu);
527
528out_unlock:
529 cpus_read_unlock();
530 mutex_unlock(&hwlat_data.lock);
531 mutex_unlock(&trace_types_lock);
532}
533
534static DECLARE_WORK(hwlat_hotplug_work, hwlat_hotplug_workfn);
535
536/*
537 * hwlat_cpu_init - CPU hotplug online callback function
538 */
539static int hwlat_cpu_init(unsigned int cpu)
540{
541 schedule_work_on(cpu, &hwlat_hotplug_work);
542 return 0;
543}
544
545/*
546 * hwlat_cpu_die - CPU hotplug offline callback function
547 */
548static int hwlat_cpu_die(unsigned int cpu)
549{
550 stop_cpu_kthread(cpu);
551 return 0;
552}
553
554static void hwlat_init_hotplug_support(void)
555{
556 int ret;
557
558 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "trace/hwlat:online",
559 hwlat_cpu_init, hwlat_cpu_die);
560 if (ret < 0)
561 pr_warn(BANNER "Error to init cpu hotplug support\n");
562
563 return;
564}
565#else /* CONFIG_HOTPLUG_CPU */
566static void hwlat_init_hotplug_support(void)
567{
568 return;
569}
570#endif /* CONFIG_HOTPLUG_CPU */
571
572/*
573 * start_per_cpu_kthreads - Kick off the hardware latency sampling/detector kthreads
574 *
575 * This starts the kernel threads that will sit on potentially all cpus and
576 * sample the CPU timestamp counter (TSC or similar) and look for potential
577 * hardware latencies.
578 */
579static int start_per_cpu_kthreads(struct trace_array *tr)
580{
581 struct cpumask *current_mask = &save_cpumask;
582 unsigned int cpu;
583 int retval;
584
585 cpus_read_lock();
586 /*
587 * Run only on CPUs in which hwlat is allowed to run.
588 */
589 cpumask_and(current_mask, cpu_online_mask, tr->tracing_cpumask);
590
591 for_each_cpu(cpu, current_mask) {
592 retval = start_cpu_kthread(cpu);
593 if (retval)
594 goto out_error;
595 }
596 cpus_read_unlock();
597
598 return 0;
599
600out_error:
601 cpus_read_unlock();
602 stop_per_cpu_kthreads();
603 return retval;
604}
605
606static void *s_mode_start(struct seq_file *s, loff_t *pos)
607{
608 int mode = *pos;
609
610 mutex_lock(&hwlat_data.lock);
611
612 if (mode >= MODE_MAX)
613 return NULL;
614
615 return pos;
616}
617
618static void *s_mode_next(struct seq_file *s, void *v, loff_t *pos)
619{
620 int mode = ++(*pos);
621
622 if (mode >= MODE_MAX)
623 return NULL;
624
625 return pos;
626}
627
628static int s_mode_show(struct seq_file *s, void *v)
629{
630 loff_t *pos = v;
631 int mode = *pos;
632
633 if (mode == hwlat_data.thread_mode)
634 seq_printf(s, "[%s]", thread_mode_str[mode]);
635 else
636 seq_printf(s, "%s", thread_mode_str[mode]);
637
638 if (mode < MODE_MAX - 1) /* if mode is any but last */
639 seq_puts(s, " ");
640
641 return 0;
642}
643
644static void s_mode_stop(struct seq_file *s, void *v)
645{
646 seq_puts(s, "\n");
647 mutex_unlock(&hwlat_data.lock);
648}
649
650static const struct seq_operations thread_mode_seq_ops = {
651 .start = s_mode_start,
652 .next = s_mode_next,
653 .show = s_mode_show,
654 .stop = s_mode_stop
655};
656
657static int hwlat_mode_open(struct inode *inode, struct file *file)
658{
659 return seq_open(file, &thread_mode_seq_ops);
660};
661
662static void hwlat_tracer_start(struct trace_array *tr);
663static void hwlat_tracer_stop(struct trace_array *tr);
664
665/**
666 * hwlat_mode_write - Write function for "mode" entry
667 * @filp: The active open file structure
668 * @ubuf: The user buffer that contains the value to write
669 * @cnt: The maximum number of bytes to write to "file"
670 * @ppos: The current position in @file
671 *
672 * This function provides a write implementation for the "mode" interface
673 * to the hardware latency detector. hwlatd has different operation modes.
674 * The "none" sets the allowed cpumask for a single hwlatd thread at the
675 * startup and lets the scheduler handle the migration. The default mode is
676 * the "round-robin" one, in which a single hwlatd thread runs, migrating
677 * among the allowed CPUs in a round-robin fashion. The "per-cpu" mode
678 * creates one hwlatd thread per allowed CPU.
679 */
680static ssize_t hwlat_mode_write(struct file *filp, const char __user *ubuf,
681 size_t cnt, loff_t *ppos)
682{
683 struct trace_array *tr = hwlat_trace;
684 const char *mode;
685 char buf[64];
686 int ret, i;
687
688 if (cnt >= sizeof(buf))
689 return -EINVAL;
690
691 if (copy_from_user(buf, ubuf, cnt))
692 return -EFAULT;
693
694 buf[cnt] = 0;
695
696 mode = strstrip(buf);
697
698 ret = -EINVAL;
699
700 /*
701 * trace_types_lock is taken to avoid concurrency on start/stop
702 * and hwlat_busy.
703 */
704 mutex_lock(&trace_types_lock);
705 if (hwlat_busy)
706 hwlat_tracer_stop(tr);
707
708 mutex_lock(&hwlat_data.lock);
709
710 for (i = 0; i < MODE_MAX; i++) {
711 if (strcmp(mode, thread_mode_str[i]) == 0) {
712 hwlat_data.thread_mode = i;
713 ret = cnt;
714 }
715 }
716
717 mutex_unlock(&hwlat_data.lock);
718
719 if (hwlat_busy)
720 hwlat_tracer_start(tr);
721 mutex_unlock(&trace_types_lock);
722
723 *ppos += cnt;
724
725
726
727 return ret;
728}
729
730/*
731 * The width parameter is read/write using the generic trace_min_max_param
732 * method. The *val is protected by the hwlat_data lock and is upper
733 * bounded by the window parameter.
734 */
735static struct trace_min_max_param hwlat_width = {
736 .lock = &hwlat_data.lock,
737 .val = &hwlat_data.sample_width,
738 .max = &hwlat_data.sample_window,
739 .min = NULL,
740};
741
742/*
743 * The window parameter is read/write using the generic trace_min_max_param
744 * method. The *val is protected by the hwlat_data lock and is lower
745 * bounded by the width parameter.
746 */
747static struct trace_min_max_param hwlat_window = {
748 .lock = &hwlat_data.lock,
749 .val = &hwlat_data.sample_window,
750 .max = NULL,
751 .min = &hwlat_data.sample_width,
752};
753
754static const struct file_operations thread_mode_fops = {
755 .open = hwlat_mode_open,
756 .read = seq_read,
757 .llseek = seq_lseek,
758 .release = seq_release,
759 .write = hwlat_mode_write
760};
761/**
762 * init_tracefs - A function to initialize the tracefs interface files
763 *
764 * This function creates entries in tracefs for "hwlat_detector".
765 * It creates the hwlat_detector directory in the tracing directory,
766 * and within that directory is the count, width and window files to
767 * change and view those values.
768 */
769static int init_tracefs(void)
770{
771 int ret;
772 struct dentry *top_dir;
773
774 ret = tracing_init_dentry();
775 if (ret)
776 return -ENOMEM;
777
778 top_dir = tracefs_create_dir("hwlat_detector", NULL);
779 if (!top_dir)
780 return -ENOMEM;
781
782 hwlat_sample_window = tracefs_create_file("window", TRACE_MODE_WRITE,
783 top_dir,
784 &hwlat_window,
785 &trace_min_max_fops);
786 if (!hwlat_sample_window)
787 goto err;
788
789 hwlat_sample_width = tracefs_create_file("width", TRACE_MODE_WRITE,
790 top_dir,
791 &hwlat_width,
792 &trace_min_max_fops);
793 if (!hwlat_sample_width)
794 goto err;
795
796 hwlat_thread_mode = trace_create_file("mode", TRACE_MODE_WRITE,
797 top_dir,
798 NULL,
799 &thread_mode_fops);
800 if (!hwlat_thread_mode)
801 goto err;
802
803 return 0;
804
805 err:
806 tracefs_remove(top_dir);
807 return -ENOMEM;
808}
809
810static void hwlat_tracer_start(struct trace_array *tr)
811{
812 int err;
813
814 if (hwlat_data.thread_mode == MODE_PER_CPU)
815 err = start_per_cpu_kthreads(tr);
816 else
817 err = start_single_kthread(tr);
818 if (err)
819 pr_err(BANNER "Cannot start hwlat kthread\n");
820}
821
822static void hwlat_tracer_stop(struct trace_array *tr)
823{
824 if (hwlat_data.thread_mode == MODE_PER_CPU)
825 stop_per_cpu_kthreads();
826 else
827 stop_single_kthread();
828}
829
830static int hwlat_tracer_init(struct trace_array *tr)
831{
832 /* Only allow one instance to enable this */
833 if (hwlat_busy)
834 return -EBUSY;
835
836 hwlat_trace = tr;
837
838 hwlat_data.count = 0;
839 tr->max_latency = 0;
840 save_tracing_thresh = tracing_thresh;
841
842 /* tracing_thresh is in nsecs, we speak in usecs */
843 if (!tracing_thresh)
844 tracing_thresh = last_tracing_thresh;
845
846 if (tracer_tracing_is_on(tr))
847 hwlat_tracer_start(tr);
848
849 hwlat_busy = true;
850
851 return 0;
852}
853
854static void hwlat_tracer_reset(struct trace_array *tr)
855{
856 hwlat_tracer_stop(tr);
857
858 /* the tracing threshold is static between runs */
859 last_tracing_thresh = tracing_thresh;
860
861 tracing_thresh = save_tracing_thresh;
862 hwlat_busy = false;
863}
864
865static struct tracer hwlat_tracer __read_mostly =
866{
867 .name = "hwlat",
868 .init = hwlat_tracer_init,
869 .reset = hwlat_tracer_reset,
870 .start = hwlat_tracer_start,
871 .stop = hwlat_tracer_stop,
872 .allow_instances = true,
873};
874
875__init static int init_hwlat_tracer(void)
876{
877 int ret;
878
879 mutex_init(&hwlat_data.lock);
880
881 ret = register_tracer(&hwlat_tracer);
882 if (ret)
883 return ret;
884
885 hwlat_init_hotplug_support();
886
887 init_tracefs();
888
889 return 0;
890}
891late_initcall(init_hwlat_tracer);
1/*
2 * trace_hwlatdetect.c - A simple Hardware Latency detector.
3 *
4 * Use this tracer to detect large system latencies induced by the behavior of
5 * certain underlying system hardware or firmware, independent of Linux itself.
6 * The code was developed originally to detect the presence of SMIs on Intel
7 * and AMD systems, although there is no dependency upon x86 herein.
8 *
9 * The classical example usage of this tracer is in detecting the presence of
10 * SMIs or System Management Interrupts on Intel and AMD systems. An SMI is a
11 * somewhat special form of hardware interrupt spawned from earlier CPU debug
12 * modes in which the (BIOS/EFI/etc.) firmware arranges for the South Bridge
13 * LPC (or other device) to generate a special interrupt under certain
14 * circumstances, for example, upon expiration of a special SMI timer device,
15 * due to certain external thermal readings, on certain I/O address accesses,
16 * and other situations. An SMI hits a special CPU pin, triggers a special
17 * SMI mode (complete with special memory map), and the OS is unaware.
18 *
19 * Although certain hardware-inducing latencies are necessary (for example,
20 * a modern system often requires an SMI handler for correct thermal control
21 * and remote management) they can wreak havoc upon any OS-level performance
22 * guarantees toward low-latency, especially when the OS is not even made
23 * aware of the presence of these interrupts. For this reason, we need a
24 * somewhat brute force mechanism to detect these interrupts. In this case,
25 * we do it by hogging all of the CPU(s) for configurable timer intervals,
26 * sampling the built-in CPU timer, looking for discontiguous readings.
27 *
28 * WARNING: This implementation necessarily introduces latencies. Therefore,
29 * you should NEVER use this tracer while running in a production
30 * environment requiring any kind of low-latency performance
31 * guarantee(s).
32 *
33 * Copyright (C) 2008-2009 Jon Masters, Red Hat, Inc. <jcm@redhat.com>
34 * Copyright (C) 2013-2016 Steven Rostedt, Red Hat, Inc. <srostedt@redhat.com>
35 *
36 * Includes useful feedback from Clark Williams <clark@redhat.com>
37 *
38 * This file is licensed under the terms of the GNU General Public
39 * License version 2. This program is licensed "as is" without any
40 * warranty of any kind, whether express or implied.
41 */
42#include <linux/kthread.h>
43#include <linux/tracefs.h>
44#include <linux/uaccess.h>
45#include <linux/cpumask.h>
46#include <linux/delay.h>
47#include "trace.h"
48
49static struct trace_array *hwlat_trace;
50
51#define U64STR_SIZE 22 /* 20 digits max */
52
53#define BANNER "hwlat_detector: "
54#define DEFAULT_SAMPLE_WINDOW 1000000 /* 1s */
55#define DEFAULT_SAMPLE_WIDTH 500000 /* 0.5s */
56#define DEFAULT_LAT_THRESHOLD 10 /* 10us */
57
58/* sampling thread*/
59static struct task_struct *hwlat_kthread;
60
61static struct dentry *hwlat_sample_width; /* sample width us */
62static struct dentry *hwlat_sample_window; /* sample window us */
63
64/* Save the previous tracing_thresh value */
65static unsigned long save_tracing_thresh;
66
67/* NMI timestamp counters */
68static u64 nmi_ts_start;
69static u64 nmi_total_ts;
70static int nmi_count;
71static int nmi_cpu;
72
73/* Tells NMIs to call back to the hwlat tracer to record timestamps */
74bool trace_hwlat_callback_enabled;
75
76/* If the user changed threshold, remember it */
77static u64 last_tracing_thresh = DEFAULT_LAT_THRESHOLD * NSEC_PER_USEC;
78
79/* Individual latency samples are stored here when detected. */
80struct hwlat_sample {
81 u64 seqnum; /* unique sequence */
82 u64 duration; /* delta */
83 u64 outer_duration; /* delta (outer loop) */
84 u64 nmi_total_ts; /* Total time spent in NMIs */
85 struct timespec timestamp; /* wall time */
86 int nmi_count; /* # NMIs during this sample */
87};
88
89/* keep the global state somewhere. */
90static struct hwlat_data {
91
92 struct mutex lock; /* protect changes */
93
94 u64 count; /* total since reset */
95
96 u64 sample_window; /* total sampling window (on+off) */
97 u64 sample_width; /* active sampling portion of window */
98
99} hwlat_data = {
100 .sample_window = DEFAULT_SAMPLE_WINDOW,
101 .sample_width = DEFAULT_SAMPLE_WIDTH,
102};
103
104static void trace_hwlat_sample(struct hwlat_sample *sample)
105{
106 struct trace_array *tr = hwlat_trace;
107 struct trace_event_call *call = &event_hwlat;
108 struct ring_buffer *buffer = tr->trace_buffer.buffer;
109 struct ring_buffer_event *event;
110 struct hwlat_entry *entry;
111 unsigned long flags;
112 int pc;
113
114 pc = preempt_count();
115 local_save_flags(flags);
116
117 event = trace_buffer_lock_reserve(buffer, TRACE_HWLAT, sizeof(*entry),
118 flags, pc);
119 if (!event)
120 return;
121 entry = ring_buffer_event_data(event);
122 entry->seqnum = sample->seqnum;
123 entry->duration = sample->duration;
124 entry->outer_duration = sample->outer_duration;
125 entry->timestamp = sample->timestamp;
126 entry->nmi_total_ts = sample->nmi_total_ts;
127 entry->nmi_count = sample->nmi_count;
128
129 if (!call_filter_check_discard(call, entry, buffer, event))
130 trace_buffer_unlock_commit_nostack(buffer, event);
131}
132
133/* Macros to encapsulate the time capturing infrastructure */
134#define time_type u64
135#define time_get() trace_clock_local()
136#define time_to_us(x) div_u64(x, 1000)
137#define time_sub(a, b) ((a) - (b))
138#define init_time(a, b) (a = b)
139#define time_u64(a) a
140
141void trace_hwlat_callback(bool enter)
142{
143 if (smp_processor_id() != nmi_cpu)
144 return;
145
146 /*
147 * Currently trace_clock_local() calls sched_clock() and the
148 * generic version is not NMI safe.
149 */
150 if (!IS_ENABLED(CONFIG_GENERIC_SCHED_CLOCK)) {
151 if (enter)
152 nmi_ts_start = time_get();
153 else
154 nmi_total_ts = time_get() - nmi_ts_start;
155 }
156
157 if (enter)
158 nmi_count++;
159}
160
161/**
162 * get_sample - sample the CPU TSC and look for likely hardware latencies
163 *
164 * Used to repeatedly capture the CPU TSC (or similar), looking for potential
165 * hardware-induced latency. Called with interrupts disabled and with
166 * hwlat_data.lock held.
167 */
168static int get_sample(void)
169{
170 struct trace_array *tr = hwlat_trace;
171 time_type start, t1, t2, last_t2;
172 s64 diff, total, last_total = 0;
173 u64 sample = 0;
174 u64 thresh = tracing_thresh;
175 u64 outer_sample = 0;
176 int ret = -1;
177
178 do_div(thresh, NSEC_PER_USEC); /* modifies interval value */
179
180 nmi_cpu = smp_processor_id();
181 nmi_total_ts = 0;
182 nmi_count = 0;
183 /* Make sure NMIs see this first */
184 barrier();
185
186 trace_hwlat_callback_enabled = true;
187
188 init_time(last_t2, 0);
189 start = time_get(); /* start timestamp */
190
191 do {
192
193 t1 = time_get(); /* we'll look for a discontinuity */
194 t2 = time_get();
195
196 if (time_u64(last_t2)) {
197 /* Check the delta from outer loop (t2 to next t1) */
198 diff = time_to_us(time_sub(t1, last_t2));
199 /* This shouldn't happen */
200 if (diff < 0) {
201 pr_err(BANNER "time running backwards\n");
202 goto out;
203 }
204 if (diff > outer_sample)
205 outer_sample = diff;
206 }
207 last_t2 = t2;
208
209 total = time_to_us(time_sub(t2, start)); /* sample width */
210
211 /* Check for possible overflows */
212 if (total < last_total) {
213 pr_err("Time total overflowed\n");
214 break;
215 }
216 last_total = total;
217
218 /* This checks the inner loop (t1 to t2) */
219 diff = time_to_us(time_sub(t2, t1)); /* current diff */
220
221 /* This shouldn't happen */
222 if (diff < 0) {
223 pr_err(BANNER "time running backwards\n");
224 goto out;
225 }
226
227 if (diff > sample)
228 sample = diff; /* only want highest value */
229
230 } while (total <= hwlat_data.sample_width);
231
232 barrier(); /* finish the above in the view for NMIs */
233 trace_hwlat_callback_enabled = false;
234 barrier(); /* Make sure nmi_total_ts is no longer updated */
235
236 ret = 0;
237
238 /* If we exceed the threshold value, we have found a hardware latency */
239 if (sample > thresh || outer_sample > thresh) {
240 struct hwlat_sample s;
241
242 ret = 1;
243
244 /* We read in microseconds */
245 if (nmi_total_ts)
246 do_div(nmi_total_ts, NSEC_PER_USEC);
247
248 hwlat_data.count++;
249 s.seqnum = hwlat_data.count;
250 s.duration = sample;
251 s.outer_duration = outer_sample;
252 s.timestamp = CURRENT_TIME;
253 s.nmi_total_ts = nmi_total_ts;
254 s.nmi_count = nmi_count;
255 trace_hwlat_sample(&s);
256
257 /* Keep a running maximum ever recorded hardware latency */
258 if (sample > tr->max_latency)
259 tr->max_latency = sample;
260 }
261
262out:
263 return ret;
264}
265
266static struct cpumask save_cpumask;
267static bool disable_migrate;
268
269static void move_to_next_cpu(bool initmask)
270{
271 static struct cpumask *current_mask;
272 int next_cpu;
273
274 if (disable_migrate)
275 return;
276
277 /* Just pick the first CPU on first iteration */
278 if (initmask) {
279 current_mask = &save_cpumask;
280 get_online_cpus();
281 cpumask_and(current_mask, cpu_online_mask, tracing_buffer_mask);
282 put_online_cpus();
283 next_cpu = cpumask_first(current_mask);
284 goto set_affinity;
285 }
286
287 /*
288 * If for some reason the user modifies the CPU affinity
289 * of this thread, than stop migrating for the duration
290 * of the current test.
291 */
292 if (!cpumask_equal(current_mask, ¤t->cpus_allowed))
293 goto disable;
294
295 get_online_cpus();
296 cpumask_and(current_mask, cpu_online_mask, tracing_buffer_mask);
297 next_cpu = cpumask_next(smp_processor_id(), current_mask);
298 put_online_cpus();
299
300 if (next_cpu >= nr_cpu_ids)
301 next_cpu = cpumask_first(current_mask);
302
303 set_affinity:
304 if (next_cpu >= nr_cpu_ids) /* Shouldn't happen! */
305 goto disable;
306
307 cpumask_clear(current_mask);
308 cpumask_set_cpu(next_cpu, current_mask);
309
310 sched_setaffinity(0, current_mask);
311 return;
312
313 disable:
314 disable_migrate = true;
315}
316
317/*
318 * kthread_fn - The CPU time sampling/hardware latency detection kernel thread
319 *
320 * Used to periodically sample the CPU TSC via a call to get_sample. We
321 * disable interrupts, which does (intentionally) introduce latency since we
322 * need to ensure nothing else might be running (and thus preempting).
323 * Obviously this should never be used in production environments.
324 *
325 * Currently this runs on which ever CPU it was scheduled on, but most
326 * real-world hardware latency situations occur across several CPUs,
327 * but we might later generalize this if we find there are any actualy
328 * systems with alternate SMI delivery or other hardware latencies.
329 */
330static int kthread_fn(void *data)
331{
332 u64 interval;
333 bool initmask = true;
334
335 while (!kthread_should_stop()) {
336
337 move_to_next_cpu(initmask);
338 initmask = false;
339
340 local_irq_disable();
341 get_sample();
342 local_irq_enable();
343
344 mutex_lock(&hwlat_data.lock);
345 interval = hwlat_data.sample_window - hwlat_data.sample_width;
346 mutex_unlock(&hwlat_data.lock);
347
348 do_div(interval, USEC_PER_MSEC); /* modifies interval value */
349
350 /* Always sleep for at least 1ms */
351 if (interval < 1)
352 interval = 1;
353
354 if (msleep_interruptible(interval))
355 break;
356 }
357
358 return 0;
359}
360
361/**
362 * start_kthread - Kick off the hardware latency sampling/detector kthread
363 *
364 * This starts the kernel thread that will sit and sample the CPU timestamp
365 * counter (TSC or similar) and look for potential hardware latencies.
366 */
367static int start_kthread(struct trace_array *tr)
368{
369 struct task_struct *kthread;
370
371 kthread = kthread_create(kthread_fn, NULL, "hwlatd");
372 if (IS_ERR(kthread)) {
373 pr_err(BANNER "could not start sampling thread\n");
374 return -ENOMEM;
375 }
376 hwlat_kthread = kthread;
377 wake_up_process(kthread);
378
379 return 0;
380}
381
382/**
383 * stop_kthread - Inform the hardware latency samping/detector kthread to stop
384 *
385 * This kicks the running hardware latency sampling/detector kernel thread and
386 * tells it to stop sampling now. Use this on unload and at system shutdown.
387 */
388static void stop_kthread(void)
389{
390 if (!hwlat_kthread)
391 return;
392 kthread_stop(hwlat_kthread);
393 hwlat_kthread = NULL;
394}
395
396/*
397 * hwlat_read - Wrapper read function for reading both window and width
398 * @filp: The active open file structure
399 * @ubuf: The userspace provided buffer to read value into
400 * @cnt: The maximum number of bytes to read
401 * @ppos: The current "file" position
402 *
403 * This function provides a generic read implementation for the global state
404 * "hwlat_data" structure filesystem entries.
405 */
406static ssize_t hwlat_read(struct file *filp, char __user *ubuf,
407 size_t cnt, loff_t *ppos)
408{
409 char buf[U64STR_SIZE];
410 u64 *entry = filp->private_data;
411 u64 val;
412 int len;
413
414 if (!entry)
415 return -EFAULT;
416
417 if (cnt > sizeof(buf))
418 cnt = sizeof(buf);
419
420 val = *entry;
421
422 len = snprintf(buf, sizeof(buf), "%llu\n", val);
423
424 return simple_read_from_buffer(ubuf, cnt, ppos, buf, len);
425}
426
427/**
428 * hwlat_width_write - Write function for "width" entry
429 * @filp: The active open file structure
430 * @ubuf: The user buffer that contains the value to write
431 * @cnt: The maximum number of bytes to write to "file"
432 * @ppos: The current position in @file
433 *
434 * This function provides a write implementation for the "width" interface
435 * to the hardware latency detector. It can be used to configure
436 * for how many us of the total window us we will actively sample for any
437 * hardware-induced latency periods. Obviously, it is not possible to
438 * sample constantly and have the system respond to a sample reader, or,
439 * worse, without having the system appear to have gone out to lunch. It
440 * is enforced that width is less that the total window size.
441 */
442static ssize_t
443hwlat_width_write(struct file *filp, const char __user *ubuf,
444 size_t cnt, loff_t *ppos)
445{
446 u64 val;
447 int err;
448
449 err = kstrtoull_from_user(ubuf, cnt, 10, &val);
450 if (err)
451 return err;
452
453 mutex_lock(&hwlat_data.lock);
454 if (val < hwlat_data.sample_window)
455 hwlat_data.sample_width = val;
456 else
457 err = -EINVAL;
458 mutex_unlock(&hwlat_data.lock);
459
460 if (err)
461 return err;
462
463 return cnt;
464}
465
466/**
467 * hwlat_window_write - Write function for "window" entry
468 * @filp: The active open file structure
469 * @ubuf: The user buffer that contains the value to write
470 * @cnt: The maximum number of bytes to write to "file"
471 * @ppos: The current position in @file
472 *
473 * This function provides a write implementation for the "window" interface
474 * to the hardware latency detetector. The window is the total time
475 * in us that will be considered one sample period. Conceptually, windows
476 * occur back-to-back and contain a sample width period during which
477 * actual sampling occurs. Can be used to write a new total window size. It
478 * is enfoced that any value written must be greater than the sample width
479 * size, or an error results.
480 */
481static ssize_t
482hwlat_window_write(struct file *filp, const char __user *ubuf,
483 size_t cnt, loff_t *ppos)
484{
485 u64 val;
486 int err;
487
488 err = kstrtoull_from_user(ubuf, cnt, 10, &val);
489 if (err)
490 return err;
491
492 mutex_lock(&hwlat_data.lock);
493 if (hwlat_data.sample_width < val)
494 hwlat_data.sample_window = val;
495 else
496 err = -EINVAL;
497 mutex_unlock(&hwlat_data.lock);
498
499 if (err)
500 return err;
501
502 return cnt;
503}
504
505static const struct file_operations width_fops = {
506 .open = tracing_open_generic,
507 .read = hwlat_read,
508 .write = hwlat_width_write,
509};
510
511static const struct file_operations window_fops = {
512 .open = tracing_open_generic,
513 .read = hwlat_read,
514 .write = hwlat_window_write,
515};
516
517/**
518 * init_tracefs - A function to initialize the tracefs interface files
519 *
520 * This function creates entries in tracefs for "hwlat_detector".
521 * It creates the hwlat_detector directory in the tracing directory,
522 * and within that directory is the count, width and window files to
523 * change and view those values.
524 */
525static int init_tracefs(void)
526{
527 struct dentry *d_tracer;
528 struct dentry *top_dir;
529
530 d_tracer = tracing_init_dentry();
531 if (IS_ERR(d_tracer))
532 return -ENOMEM;
533
534 top_dir = tracefs_create_dir("hwlat_detector", d_tracer);
535 if (!top_dir)
536 return -ENOMEM;
537
538 hwlat_sample_window = tracefs_create_file("window", 0640,
539 top_dir,
540 &hwlat_data.sample_window,
541 &window_fops);
542 if (!hwlat_sample_window)
543 goto err;
544
545 hwlat_sample_width = tracefs_create_file("width", 0644,
546 top_dir,
547 &hwlat_data.sample_width,
548 &width_fops);
549 if (!hwlat_sample_width)
550 goto err;
551
552 return 0;
553
554 err:
555 tracefs_remove_recursive(top_dir);
556 return -ENOMEM;
557}
558
559static void hwlat_tracer_start(struct trace_array *tr)
560{
561 int err;
562
563 err = start_kthread(tr);
564 if (err)
565 pr_err(BANNER "Cannot start hwlat kthread\n");
566}
567
568static void hwlat_tracer_stop(struct trace_array *tr)
569{
570 stop_kthread();
571}
572
573static bool hwlat_busy;
574
575static int hwlat_tracer_init(struct trace_array *tr)
576{
577 /* Only allow one instance to enable this */
578 if (hwlat_busy)
579 return -EBUSY;
580
581 hwlat_trace = tr;
582
583 disable_migrate = false;
584 hwlat_data.count = 0;
585 tr->max_latency = 0;
586 save_tracing_thresh = tracing_thresh;
587
588 /* tracing_thresh is in nsecs, we speak in usecs */
589 if (!tracing_thresh)
590 tracing_thresh = last_tracing_thresh;
591
592 if (tracer_tracing_is_on(tr))
593 hwlat_tracer_start(tr);
594
595 hwlat_busy = true;
596
597 return 0;
598}
599
600static void hwlat_tracer_reset(struct trace_array *tr)
601{
602 stop_kthread();
603
604 /* the tracing threshold is static between runs */
605 last_tracing_thresh = tracing_thresh;
606
607 tracing_thresh = save_tracing_thresh;
608 hwlat_busy = false;
609}
610
611static struct tracer hwlat_tracer __read_mostly =
612{
613 .name = "hwlat",
614 .init = hwlat_tracer_init,
615 .reset = hwlat_tracer_reset,
616 .start = hwlat_tracer_start,
617 .stop = hwlat_tracer_stop,
618 .allow_instances = true,
619};
620
621__init static int init_hwlat_tracer(void)
622{
623 int ret;
624
625 mutex_init(&hwlat_data.lock);
626
627 ret = register_tracer(&hwlat_tracer);
628 if (ret)
629 return ret;
630
631 init_tracefs();
632
633 return 0;
634}
635late_initcall(init_hwlat_tracer);